The Duck Test Proves ChatGPT Extracts Invisible Business Addresses from Fake Schema Code

LLM parsers ignore the rigid syntax of structured data.

ChatGPT and Perplexity extract invisible facts directly from fabricated code blocks, observes Edward Sturm.

During The Duck Test, an experiment targeting the fictional brand Ducky App T-Shirts, researchers stripped the target business address from the visible HTML body.

Operators injected the location data into a fabricated schema block.

When queried, both AI engines retrieved the hidden address from the fake code.

The system accepts this injection because the models treat structured data blocks as raw text content rather than a structural semantic layer.

This parser behavior turns schema into a valid content injection vector for LLM-based retrieval.

Practitioners can feed entity data and factual context into AI pipelines without altering the user-facing viewport.